Answers in the chat

The chat is where you read what the agent did. Your messages sit on the right and the agent's answers on the left. The tools the agent calls are not listed one by one; what they produce, such as a chart, a table, an image or a file, shows in the chat. Answers can hold charts, tables, equations, diagrams, images and files, and links get previews.

While the agent works

A line under the last message says what the agent is doing, how long the turn has run, and its speed in tokens per second, for example "searching · 12s · 48 tok/s". When sub-agents are running, it says how many instead; click it to open the agents panel.

When a turn ends, its duration and a copy button appear under the answer. Copy the response copies the whole answer.

If a turn ends with no answer, the chat says "The turn ended with no answer." with the reason Core reported when there is one, and offers Retry and Switch model. A turn that was stopped says why, for example "Stopped at your request."

Read and copy

• Select text in an answer to get a small menu: Copy, Copy message (the whole message as plain text), Copy as Markdown, and Attach as context, which quotes the selection into your next message.
• Code blocks show their language and a copy button. Shell commands also get a Run button; see Writing messages. Line numbers, wrapping and the code font size are set in Settings > Appearance.
• The rail at the side of a long chat lists your messages. Click one to jump to it; hovering shows a preview of the answer.
• Jump to the latest message appears when you scroll up.
• Back and forward arrows beside the sidebar button move between the chats you opened, like a browser.

Core stores a very long message on its own, and the app loads the full text and checks it before it shows it. If it cannot, the chat shows the start of the message and "Full answer could not be loaded." When a session's history is too long to load at once, the top of the chat says "Older history is not shown."

Links and link previews

Web addresses in answers are real links. The text drops https:// and www. and shortens long addresses; hover to see the full one.

With link previews on, each link also gets a card with the page's title, text and image. Posts on X show the author and the post, and YouTube links show the video's title, channel and thumbnail. Click a card to open the page in the app's own browser. Open in, at the top right of the card when you hover it, offers This app or Your browser.

Link previews are on by default. Turn them off in Settings > General > Chat > Link previews: "Show the title, text and image of links in chats. Pages are fetched from this computer." When the switch is off, nothing is fetched.

Previews are fetched by the app itself, never by the page you are reading. It only contacts public addresses on the standard web ports, sends no cookies, credentials or referrer, follows at most three redirects, and gives up after five seconds. Results are kept in memory for a few hours.

A link in the text opens in your default browser. To open these in the app's own browser instead, check Open chat links here in the browser's options; see Browser and terminal.

Files the agent mentions

A file path in an answer becomes a link showing the file name, when that file exists in the session's folder. An image or video opens in the viewer. Any other file is only revealed in Finder or your file manager; the app never opens it for you. Paths to missing files, or files outside the folder, stay plain text.

Images and videos

Images and videos the agent generates appear in the chat and stay there after a reload, a restart or a switch to another session. Several images show as a viewer with thumbnails. Videos play inline, muted, with the usual controls.

Click an image or video to open it large; Escape closes it. Save a copy saves the file where you choose, and Show in Finder (Show in folder on Windows and Linux) reveals it.

Charts, tables, images and files from tools

When a tool returns a chart, a table, an image or a file, the app shows it in the chat where the tool ran, and it survives a reload:

• Charts are drawn as described in the next section.
• Tables are real tables, with Copy CSV.
• Images open in the viewer.
• Files show as a card with the file's name, size and type, and Show in Finder (Show in folder on Windows and Linux).

The app checks each item against what Core recorded before it shows it. If an item cannot be read or has changed, a short notice such as "Attachment unavailable" takes its place.

Charts

When the agent answers with data, it can draw a chart instead of text art.

• Bar, line, area and scatter charts, pies and donuts, stacked bars and horizontal bars, with up to 12 series and 5,000 points.
• Hover or click a mark to read its exact values. With the keyboard, Tab enters the chart and the arrow keys move between points.
• Click a series in the legend to hide or show it.
• Every chart has a data table and Copy CSV.

Price charts get a finance layout: candles or OHLC bars, line and area series, moving averages, markers, and a separate volume band. The header shows the last price, the change and the "As of" date, and a legend follows the crosshair. Range buttons from one month to all of the data change the span. View data, Copy data and Save image sit in the header.

A chart that cannot be drawn shows why, with its source.

Equations and diagrams

• Equations written in TeX render as math, inline or on their own line. A displayed equation has its source and a copy control. Invalid TeX shows as its source.
• Diagrams: flowcharts from a small subset of Mermaid (flowchart or graph, with rectangles, rounded boxes, diamonds and labeled arrows), or from a simple node and edge format. A diagram first fits the message; Fit, 100% and the zoom buttons go up to 300%. Each has a text description and its source.

These formats never run code: HTML, SVG and JavaScript blocks are shown as code, and nothing is fetched from the network to draw them.

Session context

The session context panel shows how full the model's context window is and what the session has used. Open it from the project name at the top of the chat, then Session context. It shows:

• the selected model, the share of its context window in use, and whether that is measured or estimated;
• estimated sources: messages, system instructions, system tools, loaded MCP tools and memory files;
• the last completed turn and the session total: tokens, input and output, and estimated cost;
• usage by model, since switching models does not reset the totals, and the cache hit rate when the provider reports it.

While a turn runs, the panel refreshes every five seconds. Refresh reads a new snapshot. The panel can float over the chat or be pinned to the right side. Token activity opens your usage history in Settings > Profile.